Google Advances Responsible AI with Gemma 2 Models

In a significant leap towards responsible AI, Google has unveiled Gemma 2, their latest suite of open models, available in 27 billion (27B) and 9 billion (9B) parameter sizes. The 27B model quickly earned a top spot on the LMSYS Chatbot Arena leaderboard, outperforming several larger models in real-world conversations.

Gemma 2’s launch is part of Google’s ongoing commitment to developing AI that is not only high-performing but also safe and accessible. To further this mission, Google introduced three new additions to the Gemma 2 family:

New Additions to the Gemma 2 Family

• Gemma 2 2B: A new iteration of the popular 2 billion (2B) parameter model, offering built-in safety enhancements and a balance of performance and efficiency.

• ShieldGemma: A suite of safety content classifier models built on Gemma 2, designed to filter AI inputs and outputs, ensuring user safety.

• Gemma Scope: A model interpretability tool providing deep insights into the inner workings of Gemma 2 models.

These additions are designed to help researchers and developers create safer user experiences, gain unprecedented insights into AI models, and confidently deploy powerful AI on various devices, driving innovation forward.

Introducing Gemma 2 2B

The Gemma 2 2B model is highly anticipated for its ability to deliver exceptional results by learning from larger models through distillation. It has already surpassed all GPT-3.5 models on the Chatbot Arena, showcasing its exceptional conversational AI capabilities.

Key features of Gemma 2 2B include:

• Exceptional performance: Best-in-class performance for its size, surpassing other models in its category.

• Flexible and cost-effective deployment: Runs efficiently on a range of hardware, from edge devices and laptops to cloud deployments with Vertex AI and Google Kubernetes Engine (GKE). Optimized with the NVIDIA TensorRT-LLM library, it is available as an NVIDIA NIM, suitable for data centers, local workstations, PCs, and edge devices using NVIDIA RTX, GeForce RTX GPUs, or Jetson modules. It integrates seamlessly with Keras, JAX, Hugging Face, NVIDIA NeMo, Ollama, Gemma.cpp, and soon MediaPipe.

• Open and accessible: Available under commercially-friendly terms for research and commercial use, small enough to run on the free tier of T4 GPUs in Google Colab.

Starting today, Gemma 2’s model weights can be downloaded from Kaggle, Hugging Face, and Vertex AI Model Garden, or explored in Google AI Studio.

Introducing ShieldGemma

Deploying responsible AI models requires substantial effort. To assist developers, Google presents ShieldGemma, a series of advanced safety classifiers designed to detect and mitigate harmful content in AI model inputs and outputs. ShieldGemma focuses on:

• Hate speech

• Harassment

• Sexually explicit content

• Dangerous content

These classifiers complement Google’s Responsible AI Toolkit and existing Google Cloud classifiers, offering flexible sizes for various needs and leveraging NVIDIA speed optimizations.

Gemma Scope: Unveiling AI Insights

Gemma Scope provides unprecedented transparency into the decision-making processes of Gemma 2 models. Using sparse autoencoders (SAEs), it acts like a powerful microscope, offering detailed insights into how Gemma 2 identifies patterns, processes information, and makes predictions.

Key features of Gemma Scope include:

• Open SAEs: Over 400 freely available SAEs covering all layers of Gemma 2 2B and 9B.

• Interactive demos: Explore SAE features and analyze model behavior on Neuronpedia.

• Easy-to-use repository: Access code and examples for interfacing with SAEs and Gemma 2.

Learn more about Gemma Scope on the Google DeepMind blog, technical report, and developer documentation.

Advancing Towards Responsible AI

These releases reflect Google’s ongoing commitment to equipping the AI community with the tools and resources needed to build a future where AI benefits everyone. Google believes in open access, transparency, and collaboration to develop safe and beneficial AI.

Join the Journey Towards Responsible AI

Experience the power of Gemma 2 2B by downloading it or trying it with NVIDIA NIM or Google AI Studio. Explore ShieldGemma to build safer AI applications and uncover the inner workings of Gemma 2 with Gemma Scope on Neuronpedia. Together, we can advance towards a more responsible and beneficial AI future.

As we embrace a new era of on-device, local AI, lightweight and efficient models are crucial for running AI directly on phones and laptops. With Gemma 2 outperforming GPT-3.5 Turbo at just a fraction of its size, Google is not only showcasing what's possible but also establishing its leadership in the small model space.